Google

Data with no benefit should be left out for computational reasons. In the cat-detector example, scanned historical documents that contain nothing resembling a cat and look completely unlike the dev/test distribution have negligible benefit, and including them would waste computation resources and neural-network representation capacity.

Adding More Training Data Does Not Always Help

Machine Learning Yearning, a free ebook from Andrew Ng, teaches you how to structure Machine Learning projects.
https://www.deeplearning.ai/machine-learning-yearning/

Note: The content of the book is aligned with the Coursera Deeplearning.ai specialization.  https://www.deeplearning.ai/deep-learning-specialization/ 

Machine Learning Yearning (Deeplearning.ai)

In the cat-detector example, why does Machine Learning Yearning recommend excluding scanned historical documents that look nothing like the dev/test distribution?

True or False: Adding more training data always improves model performance when training, dev, and test sets share the same distribution.

According to Machine Learning Yearning, if the dev error curve has _____ (i.e., flattened out), adding more training data will not help you reach your performance goal.

Why does Machine Learning Yearning recommend leaving out training data that has no benefit for your model?

True or False: According to Machine Learning Yearning, adding more training data can actually hurt model performance.

If the dev error curve has _____, you can immediately tell that adding more training data won't help reach your performance goal.

Match each concept to its correct description in Machine Learning Yearning's discussion of when adding training data does not help.

Order the steps for using a learning curve to decide whether to collect more training data, as described in Machine Learning Yearning.

In the cat-detector example, why should a large collection of scanned historical documents be excluded from training?

True or False: According to Machine Learning Yearning, inspecting the learning curve can prevent wasting months collecting data that turns out not to help.

According to Machine Learning Yearning, data that has no _____ should be left out of training for computational reasons.

Match each training data scenario to the recommended action from Machine Learning Yearning.

Order the reasoning steps for evaluating whether a new data source (e.g., scanned documents) should be added to training, per Machine Learning Yearning.

Question:
Explain how incorporating training data that has no benefit impacts a machine learning system's training process, specifically focusing on computational resources and neural-network representation capacity. Ground your answer in the concept of excluding data that deviates entirely from the dev/test distribution.

Sample answer:
According to the source, adding training data that has no benefit should be left out for computational reasons. Including irrelevant data wastes valuable computational resources during training. Furthermore, including such data wastes the neural-network's representation capacity, as the network dedicates its parameters to learning and representing features that are irrelevant to the dev/test distribution.

Key points:
- Data with no benefit should be left out of training for computational reasons.
- Including irrelevant data wastes neural-network representation capacity.
- Irrelevant data forces the network to learn features that do not align with the dev/test distribution.

Rubric:
A satisfactory response must mention that unhelpful training data consumes/wastes computational resources during training and also wastes the neural-network's representation capacity by forcing it to represent patterns that do not benefit the dev/test distribution.

Analyzing Computational and Representational Costs of Unhelpful Training Data

Case context:
You are building a casual cat-detector system where your target dev/test sets consist of everyday photographs of people, places, landmarks, and animals. You are offered a massive dataset of scanned historical text documents containing no cats or cat-like features to add to your training set.

Question:
Based on the principles in Machine Learning Yearning, decide whether you should include these scanned historical documents in your training data, and explain the consequences of doing so regarding model capacity and computational efficiency.

Sample answer:
The scanned historical documents should be excluded from the training data. Because they look completely unlike the dev/test distribution and contain nothing resembling a cat, they offer negligible benefit. Including them would waste computational resources during training and consume neural-network representation capacity on features irrelevant to the target task.

Key points:
- Exclude the scanned historical documents from the training data.
- The scanned documents look completely unlike the dev/test distribution and have negligible benefit.
- Including them wastes computational resources and neural-network representation capacity.

Rubric:
The student should state that the documents must be excluded. They must identify that inclusion leads to wasted computational resources and wasted neural-network representation capacity on irrelevant features.

Evaluating the Inclusion of Historical Document Scans in a Casual Cat-Detector Model

Question:
Besides wasting computational resources, what other negative impact does including unbeneficial training data (like scanned text in a casual photo detector) have on a neural network?

Sample answer:
Including unbeneficial training data wastes the neural-network's representation capacity by forcing it to represent features that do not help with the target dev/test distribution.

Key points:
- Wastes neural-network representation capacity.
- Forces the network to allocate parameters/capacity to irrelevant features.

Rubric:
The response must state that it wastes the representation capacity of the neural network.

Learn Before

Related